Skip to content

agent2model

Turn your LangGraph agent into a small open model that runs with no orchestrator — near-frontier quality at a fraction of the inference cost.

agent2model (PyPI: agent2model) takes a procedural agent workflow — written as YAML or imported straight from a LangGraph StateGraph — and bakes the whole procedure into a small model's weights via synthetic-data fine-tuning. The result self-orchestrates at runtime: there is no external orchestrator and no per-turn frontier call.

Unlike prompt-optimizers (DSPy, GEPA) that keep a runtime program, and unlike agent frameworks (LangGraph, CrewAI) that run the procedure live every turn, agent2model removes the orchestrator entirely — the procedure lives in the weights.

Based on Dennis et al. 2026, Compiling Agentic Workflows into LLM Weights (arXiv:2605.22502), which reports near-frontier quality at 128–462× lower inference cost. Those are the paper's figures; this repo's own independently-reproduced numbers are tracked in the benchmarks and are still being filled in.

The idea

Most agent frameworks inject the procedure into a frontier model's prompt on every turn. That is expensive and slow. Instead, agent2model:

  1. takes your procedure as a Flowchart IR (YAML, or imported from LangGraph);
  2. generates synthetic conversations that walk the procedure, via Claude;
  3. fine-tunes a small open model (Qwen 2.5 3B / Qwen3 8B) on those conversations — full-parameter SFT, no LoRA;
  4. evaluates the result against frontier baselines on the paper's 5-criterion rubric, and serves it behind an OpenAI-compatible endpoint.

The flowchart structure never appears in the training data — the model learns to run the procedure from natural dialogue alone.

The four-command journey

agent2model compile examples/travel_booking/flowchart.yaml --out build/travel
agent2model generate build/travel --n 2000 --model claude-sonnet-4-5
agent2model train    build/travel --base Qwen/Qwen2.5-3B-Instruct --size 3b --epochs 20
agent2model eval     build/travel --baselines in_context,langgraph --n 200
# or: agent2model serve build/travel --port 8000

See the Quickstart to run it end to end (including with no local GPU via Modal).

Where to go next

Scope

v1 ships full-parameter SFT only (no LoRA — the paper's companion shows it fails on procedural tasks), single-agent procedural workflows, and cloud recipes. RLHF/DPO, online learning, tool use during inference, and multi-agent handoffs are v2+. License: Apache-2.0.